X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;ds=sidebyside;f=src%2Fnnue%2Fevaluate_nnue.cpp;h=5416f13e1f77b502c255a22792072605832a32e3;hb=7ffae17f85709e49672a0e98e136b66aea067b2c;hp=ed1388812e009874ab26fc896e92258374eea2a7;hpb=fc27d158c012341593518a05abf51903ecbcb495;p=stockfish

diff --git a/src/nnue/evaluate_nnue.cpp b/src/nnue/evaluate_nnue.cpp
index ed138881..5416f13e 100644
--- a/src/nnue/evaluate_nnue.cpp
+++ b/src/nnue/evaluate_nnue.cpp
@@ -1,6 +1,6 @@
 /*
   Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2020 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2021 The Stockfish developers (see AUTHORS file)
 
   Stockfish is free software: you can redistribute it and/or modify
   it under the terms of the GNU General Public License as published by
@@ -25,34 +25,14 @@
 #include "../position.h"
 #include "../misc.h"
 #include "../uci.h"
+#include "../types.h"
 
 #include "evaluate_nnue.h"
 
-namespace Eval::NNUE {
-
-  uint32_t kpp_board_index[PIECE_NB][COLOR_NB] = {
-   // convention: W - us, B - them
-   // viewed from other side, W and B are reversed
-      { PS_NONE,     PS_NONE     },
-      { PS_W_PAWN,   PS_B_PAWN   },
-      { PS_W_KNIGHT, PS_B_KNIGHT },
-      { PS_W_BISHOP, PS_B_BISHOP },
-      { PS_W_ROOK,   PS_B_ROOK   },
-      { PS_W_QUEEN,  PS_B_QUEEN  },
-      { PS_W_KING,   PS_B_KING   },
-      { PS_NONE,     PS_NONE     },
-      { PS_NONE,     PS_NONE     },
-      { PS_B_PAWN,   PS_W_PAWN   },
-      { PS_B_KNIGHT, PS_W_KNIGHT },
-      { PS_B_BISHOP, PS_W_BISHOP },
-      { PS_B_ROOK,   PS_W_ROOK   },
-      { PS_B_QUEEN,  PS_W_QUEEN  },
-      { PS_B_KING,   PS_W_KING   },
-      { PS_NONE,     PS_NONE     }
-  };
+namespace Stockfish::Eval::NNUE {
 
   // Input feature converter
-  AlignedPtr<FeatureTransformer> feature_transformer;
+  LargePagePtr<FeatureTransformer> feature_transformer;
 
   // Evaluation function
   AlignedPtr<Network> network;
@@ -70,14 +50,22 @@ namespace Eval::NNUE {
     std::memset(pointer.get(), 0, sizeof(T));
   }
 
+  template <typename T>
+  void Initialize(LargePagePtr<T>& pointer) {
+
+    static_assert(alignof(T) <= 4096, "aligned_large_pages_alloc() may fail for such a big alignment requirement of T");
+    pointer.reset(reinterpret_cast<T*>(aligned_large_pages_alloc(sizeof(T))));
+    std::memset(pointer.get(), 0, sizeof(T));
+  }
+
   // Read evaluation function parameters
   template <typename T>
-  bool ReadParameters(std::istream& stream, const AlignedPtr<T>& pointer) {
+  bool ReadParameters(std::istream& stream, T& reference) {
 
     std::uint32_t header;
     header = read_little_endian<std::uint32_t>(stream);
     if (!stream || header != T::GetHashValue()) return false;
-    return pointer->ReadParameters(stream);
+    return reference.ReadParameters(stream);
   }
 
   }  // namespace Detail
@@ -110,29 +98,47 @@ namespace Eval::NNUE {
     std::string architecture;
     if (!ReadHeader(stream, &hash_value, &architecture)) return false;
     if (hash_value != kHashValue) return false;
-    if (!Detail::ReadParameters(stream, feature_transformer)) return false;
-    if (!Detail::ReadParameters(stream, network)) return false;
+    if (!Detail::ReadParameters(stream, *feature_transformer)) return false;
+    if (!Detail::ReadParameters(stream, *network)) return false;
     return stream && stream.peek() == std::ios::traits_type::eof();
   }
 
   // Evaluation function. Perform differential calculation.
   Value evaluate(const Position& pos) {
 
-    alignas(kCacheLineSize) TransformedFeatureType
-        transformed_features[FeatureTransformer::kBufferSize];
+    // We manually align the arrays on the stack because with gcc < 9.3
+    // overaligning stack variables with alignas() doesn't work correctly.
+
+    constexpr uint64_t alignment = kCacheLineSize;
+
+#if defined(ALIGNAS_ON_STACK_VARIABLES_BROKEN)
+    TransformedFeatureType transformed_features_unaligned[
+      FeatureTransformer::kBufferSize + alignment / sizeof(TransformedFeatureType)];
+    char buffer_unaligned[Network::kBufferSize + alignment];
+
+    auto* transformed_features = align_ptr_up<alignment>(&transformed_features_unaligned[0]);
+    auto* buffer = align_ptr_up<alignment>(&buffer_unaligned[0]);
+#else
+    alignas(alignment)
+      TransformedFeatureType transformed_features[FeatureTransformer::kBufferSize];
+    alignas(alignment) char buffer[Network::kBufferSize];
+#endif
+
+    ASSERT_ALIGNED(transformed_features, alignment);
+    ASSERT_ALIGNED(buffer, alignment);
+
     feature_transformer->Transform(pos, transformed_features);
-    alignas(kCacheLineSize) char buffer[Network::kBufferSize];
     const auto output = network->Propagate(transformed_features, buffer);
 
     return static_cast<Value>(output[0] / FV_SCALE);
   }
 
   // Load eval, from a file stream or a memory stream
-  bool load_eval(std::string streamName, std::istream& stream) {
+  bool load_eval(std::string name, std::istream& stream) {
 
     Initialize();
-    fileName = streamName;
+    fileName = name;
     return ReadParameters(stream);
   }
 
-} // namespace Eval::NNUE
+} // namespace Stockfish::Eval::NNUE